Answers are grounded in this paper and cite their sources.
This paper studies compact reinforcement learning for scheduling precedence-constrained directed acyclic graph (DAG) workloads across heterogeneous cloud nodes. A graph-conditioned tabular Q-learning controller selects among four transparent energy–finish-time dispatch rules, using a temporal-difference update with discounted next-state value. An always-on active/idle server-power model accounts for both execution and waiting intervals. The scheduler is evaluated on 120 held-out synthetic workflows after 800 training episodes against random, active-energy-only, greedy modeled-energy, earliest-finish-time, fixed-weight, and resource-state-only Q-learning comparators. Graph-Q yields 22717.4 J mean modeled energy and 68.48 s mean makespan; relative to fixed weighting, it reduces makespan by 5.31% while increasing modeled energy by 1.20%. Paired confidence intervals quantify these differences. The resource-state-only ablation achieves comparable modeled energy and slightly shorter completion time, delineating the value of compact graph features within this controlled simulation.
No takes yet. Share an insight, caveat, or question.
Felix Neumann (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: